[Feature] Add C++ implementation for bbox_overlaps (open-mmlab#2477)

* add ops bbox_overlaps

* format code

* Return the pytorch version

* Intermediate modification

* Solve problems in parameter passing

* revise bug

* "add test case"

mmcv/ops/bbox.py

@@ -106,25 +106,17 @@ def bbox_overlaps(bboxes1: torch.Tensor,
 
     rows = bboxes1.size(0)
     cols = bboxes2.size(0)
+
     if aligned:
         assert rows == cols
+        ious = bboxes1.new_zeros(rows)
+    else:
+        ious = bboxes1.new_zeros((rows, cols))
 
     if rows * cols == 0:
-        return bboxes1.new(rows, 1) if aligned else bboxes1.new(rows, cols)
-
-    if bboxes1.device.type == 'cpu':
-        return _bbox_overlaps_cpu(
-            bboxes1, bboxes2, mode=mode, aligned=aligned, offset=offset)
-    else:
-        if aligned:
-            ious = bboxes1.new_zeros(rows)
-        else:
-            ious = bboxes1.new_zeros((rows, cols))
-        ext_module.bbox_overlaps(
-            bboxes1,
-            bboxes2,
-            ious,
-            mode=mode_flag,
-            aligned=aligned,
-            offset=offset)
         return ious
+
+    ext_module.bbox_overlaps(
+        bboxes1, bboxes2, ious, mode=mode_flag, aligned=aligned, offset=offset)
+
+    return ious

mmcv/ops/csrc/pytorch/cpu/bbox_overlaps_cpu.cpp

@@ -0,0 +1,65 @@
+// Copyright(c) OpenMMLab.All rights reserved.
+#include "pytorch_cpp_helper.hpp"
+#include "pytorch_device_registry.hpp"
+
+using torch::indexing::None;
+using torch::indexing::Slice;
+
+void bbox_overlaps_cpu_kernel(const Tensor boxes1, const Tensor boxes2,
+                              Tensor ious, const int mode_flag,
+                              const bool aligned, const int offset) {
+  Tensor temp_ious;
+  if (aligned) {
+    Tensor lt = torch::max(boxes1.index({Slice(None), Slice({None, 2})}),
+                           boxes2.index({Slice(None), Slice({None, 2})}));
+    Tensor rb = torch::min(boxes1.index({Slice(None), Slice(2)}),
+                           boxes2.index({Slice(None), Slice(2)}));
+    Tensor wh = (rb - lt + offset).clamp(0.f, INT_MAX * 1.f);
+    Tensor overlap = wh.index({Slice(None), 0}) * wh.index({Slice(None), 1});
+    Tensor area1 = (boxes1.index({Slice(None), 2}) -
+                    boxes1.index({Slice(None), 0}) + offset) *
+                   (boxes1.index({Slice(None), 3}) -
+                    boxes1.index({Slice(None), 1}) + offset);
+    if (mode_flag == 0) {
+      Tensor area2 = (boxes2.index({Slice(None), 2}) -
+                      boxes2.index({Slice(None), 0}) + offset) *
+                     (boxes2.index({Slice(None), 3}) -
+                      boxes2.index({Slice(None), 1}) + offset);
+      temp_ious = overlap / (area1 + area2 - overlap);
+    } else {
+      temp_ious = overlap / area1;
+    }
+  } else {
+    Tensor lt = torch::max(boxes1.index({Slice(None), None, Slice({None, 2})}),
+                           boxes2.index({Slice(None), Slice({None, 2})}));
+    Tensor rb = torch::min(boxes1.index({Slice(None), None, Slice(2)}),
+                           boxes2.index({Slice(None), Slice(2)}));
+    Tensor wh = (rb - lt + offset).clamp(0.f, INT_MAX * 1.f);
+    Tensor overlap = wh.index({"...", 0}) * wh.index({"...", 1});
+    Tensor area1 = (boxes1.index({Slice(None), 2}) -
+                    boxes1.index({Slice(None), 0}) + offset) *
+                   (boxes1.index({Slice(None), 3}) -
+                    boxes1.index({Slice(None), 1}) + offset);
+    if (mode_flag == 0) {
+      Tensor area2 = (boxes2.index({Slice(None), 2}) -
+                      boxes2.index({Slice(None), 0}) + offset) *
+                     (boxes2.index({Slice(None), 3}) -
+                      boxes2.index({Slice(None), 1}) + offset);
+      temp_ious =
+          overlap / (area1.index({Slice(None), None}) + area2 - overlap);
+    } else {
+      temp_ious = overlap / area1.index({Slice(None), None});
+    }
+  }
+  ious.copy_(temp_ious);
+}
+
+void bbox_overlaps_cpu(const Tensor boxes1, const Tensor boxes2, Tensor ious,
+                       const int mode, const bool aligned, const int offset) {
+  bbox_overlaps_cpu_kernel(boxes1, boxes2, ious, mode, aligned, offset);
+}
+
+void bbox_overlaps_impl(const Tensor boxes1, const Tensor boxes2, Tensor ious,
+                        const int mode, const bool aligned, const int offset);
+
+REGISTER_DEVICE_IMPL(bbox_overlaps_impl, CPU, bbox_overlaps_cpu);

tests/test_ops/test_bbox.py

@@ -35,6 +35,14 @@ def _test_bbox_overlaps(self, device='cpu', dtype=torch.float):
         out = bbox_overlaps(b1, b2, offset=1)
         assert np.allclose(out.cpu().numpy(), should_output, 1e-2)
 
+        b1 = torch.tensor([[10.0 + i, 10.0 + i, 30.0 + i, 30.0 + i]
+                           for i in range(1000)]).to(device).type(dtype)
+        b2 = torch.tensor([[20.0 + i, 20.0 + i, 40.0 + i, 40.0 + i]
+                           for i in range(1000)]).to(device).type(dtype)
+        should_output = np.array([1 / 7] * 1000)
+        out = bbox_overlaps(b1, b2, aligned=True)
+        assert np.allclose(out.cpu().numpy(), should_output, 1e-2)
+
     @pytest.mark.parametrize('device', [
         'cpu',
         pytest.param(